Subagent swarms

Orchestrated subagents and swarm-style workflow in Codex:

Bug hunt swarm:

…fire 4 sub agents, and use main as an orchestrator in order to hunt for the precise root cause and hopefully resolution solution(s) for a given bug

Codex Subagents:

Codex can run subagent workflows by spawning specialized agents in parallel and then collecting their results in one response. This can be particularly helpful for complex tasks that are highly parallel, such as codebase exploration or implementing a multi-step feature plan.

Custom subagents:

…create your own custom subagents ~/.codex/agents/bonsai.toml

The zeitgeist:

subagents maxxing

Voice workflows

Appointments and intakes with gpt-realtime-1.5:

We built a clinic concierge demo for a Singapore health clinic with gpt-realtime-1.5. It speaks naturally with patients, collects the right details, and books appointments in real time.

Voice-to-frontend with Codex:

Using voice transcription with Codex Spark in the pop out window is wild for rapid front-end development! You don't even have time to use steering

Codex as Chief of Staff

For knowledge work:

Plugins enable Codex to do most of my knowledge work. It reads and responds to emails and Slack. It builds models in Sheets and handles all of my note taking for me. Codex is quickly becoming my assistant. It understands me and what I’m doing enough to help get real work done.

For dinner planning:

Our incredible comms leader, @lindsmccallum, planned a closed door dinner for the Codex team in a fraction of the time it would normally take - thanks to Codex.

She used the Codex App to:

  • compile the invite list
  • send out invitations
  • hourly scan of her emails to update RSVP status
  • populate a doc with bios on every attendee
  • create a mini app to plan the seating chart

Inbox triage:

Getting codex's help to me climb from under my inbox mountain

Automation density:

I've been using plug-ins a ton internally. I have about 58 automations and 30 plug-ins and I've automated everything except the part where I have to come up with ideas and talk to people

The Infinity Machine

From Sebastian Mallaby:

Even by the standard of a tech industry stacked with so-called geniuses, Demis Hassabis is a special case. Born poor in North London to immigrant parents, a chess prodigy by age five and wizard coder in his teens, he turned down a seven figure offer before turning 18 to feed his insatiable scientific curiosity at Cambridge. Later, he added a neuroscience PhD to his computer science skills to pursue the dream of artificial general intelligence, the ultimate goal being to unravel the mysteries of biology and theoretical physics and to usher in super-abundance. Alongside a small group of fellow travelers, that is the path he is still on, leading the AI research at Google, winning a Nobel Prize along the way, and imagining machines that will compound, or possibly supplant, the human understanding of the universe.

AI compute in orbit

Space as a new compute layer. From Starcloud:

The round comes after the successful deployment of our first satellite, Starcould-1, a few months ago, which had the first @NVIDIA H100 on board and was the first to train an LLM in space. The funds will be used to develop our third satellite, which aims to be cost-competitive with Earth-based data centers in terms of AI inference cost.

Alignment priors wash out under RL

Reading moral books doesn’t matter much if the later training regime teaches different habits.

From Tomek Korbak and team at OpenAI:

Can midtraining on docs about aligned AI bake in alignment priors for agents? We report an experiment where those priors are quickly washed away by RL and fail to generalize to agentic settings. But that cuts both ways: priors that AIs are misaligned fade too!